Everything

The Rounds scoops up $24M to bring its ‘household restocking’ delivery service to more markets

The Rounds, the startup that delivers recurring grocery and household essentials in reusable packaging, announced on Monday its $24 million Series B funding round. The new capital will go toward product development, hiring and expanding the service to additional markets.

Alongside the funding announcement, The Rounds introduced new features for customers, including the ability to make one-time purchases and the flexibility of choosing their delivery day. It also enhanced its AI-powered algorithm to provide smarter predictions on when customer’s products need to be restocked. 

Launched in 2019, The Rounds offers a $10 subscription for shoppers to order weekly bundles, ranging from things like household basics and personal care items to pantry staples and dry goods, as well as items from local bakeries, coffee roasters or other sellers. The Rounds recently added more than 100 new products, bringing more fresh produce options and home essentials to its catalog. The new additions bring the new total to 12,000 SKUs, almost double the amount available last year when it was at 650 SKUs. 

Users can customize how frequently they’d like certain items restocked, setting up regularly scheduled deliveries designed so customers never run out of necessary supplies like toilet paper, milk and eggs. Until recently, customers couldn’t choose when the weekly “Refill Day” was as The Rounds determined the date for each neighborhood. Now, they have the flexibility to select their own refill day to better suit their schedule.

Committed to sustainability, The Rounds delivers items in reusable bags and refillable containers. The company also uses what it calls a “reverse logistics” system where the customers can leave their empty containers on their doorsteps for delivery workers to return, sanitize and recirculate — significantly reducing waste. To date, the company has saved more than 1 million pounds of packaging waste, a significant milestone compared to the 375,000 pounds of waste TechCrunch previously reported. 

Image Credits: The Rounds

The Rounds recently introduced a new feature that allows users to make one-time purchases instead of automatically restocking items before the next delivery. This helps prevent users from receiving products they don’t need. Users have always been able to add or remove items before delivery; this new option ensures that they won’t forget to remove unwanted items.

The company applies machine learning to provide smarter predictions on when everyday items need to be refilled and to offer personalized product recommendations. According to co-founder and CEO Alex Torrey, these predictions are even more accurate than before, bringing the AI closer to being a “psychic home manager” that knows exactly what the customer needs to restock and when.

“The accuracy of the prediction, which is where our sustainable approach of using reusable packaging has an extremely valuable benefit,” Torrey said. “By collecting empty containers when you finish, we actually create a reinforcement learning feedback loop for our prediction engine that allows us to more accurately and faster predict your consumption at your household level. [All that competitors] can do is make predictions based on purchase history. We can make predictions based on purchase history [but] we also have that data based on consumption.” 

He added that, at some point in the future, the goal is to get its prediction algorithm to be so advanced that, for example, “pollen count influences our prediction.” Torrey explained, “I view a world where we’re able to say, ‘Hey, we know you have allergies, the pollen count is up, or the pollen count appears to be going up next week. We’re going to accelerate your box of tissues.’”

Image Credits: The Rounds

The Rounds is currently available in select U.S. markets, such as Atlanta, D.C. and Philadelphia. However, in early 2025, The Rounds will expand to “a couple” more markets, Torrey told us, yet declined to say where. What we do know, though, is that New York City is not on the company’s radar yet. 

“The reason we picked our markets is to really help us learn as much as possible. The downfall of launching in New York is it doesn’t teach you how to operate anywhere other than New York… Atlanta is a lot like Houston or Los Angeles. Philly is a lot like Boston. So we’ve been thinking about markets where we can learn a lot and transfer those learnings to make sure that we’re a company that works everywhere and can have a really big impact by serving as many members as possible,” Torrey noted. 

The recent round was led by Moderne Ventures, with participation from Construct Capital, First Round Capital, FJ Labs and Redpoint Ventures, among others. The company is currently focused on expanding its machine learning and data science teams and is looking to bring in a chief marketing officer and chief operations officer.

To date, The Rounds has raised $66 million in funding.

Read More

Eppo receives new funding to expand its app, website, and AI experimentation business

The AI industry continues to release tons of new models, and companies looking to stay competitive are racing to adopt them for their purposes. In fact, nearly 10% of businesses plan to spend a whopping $25 million this year on AI initiatives, according to tech consulting firm Searce.

But while lots of money is being spent on AI, it’s unclear the ROI is there. Half of all AI leaders aren’t sure how to calculate or demonstrate the value of AI projects, according to Gartner.

Ex-Airbnb data scientist Chetan Sharma makes the case that figuring out AI ROI is no heavy lift with the right tooling. Sharma is the co-founder of Eppo, an experimentation platform that lets customers evaluate and customize AI models for specific use cases. Beyond its model evaluation suite, Eppo provides a general A/B testing platform and service for apps and websites.

“With new AI models launching weekly and companies pouring millions into them, A/B testing offers a cost-effective way to evaluate their effectiveness without overspending,” Sharma told TechCrunch. “Eppo helps companies identify which models truly deliver value and enables smarter, more sustainable decisions in an environment of rapid innovation and escalating costs.”

Eppo competes with a number of experimentation and A/B testing startups in the market, including Split, Statsig and Optimizely. Big tech giants like AWS, Microsoft Azure and Google Cloud also offer a growing number of model fine-tuning and evaluation tools.

But Sharma says that Eppo stands apart from the crowd thanks to features like its “contextual bandit” system, which automatically spots new variants of customers’ websites, apps or AI models and actively explores the performance of those variants by serving increasing load or traffic to them.

Eppo’s back-end dashboard.Image Credits: Eppo

“Experimentation drives velocity and accelerates growth by stripping away bureaucratic — and often incorrect — decisions by committee while tightly tethering initiatives to growth metrics, killing bad ideas fast while canonizing good ideas for reinvestment,” Sharma said. “Eppo’s approach to live ‘online-eval’ tests of AI models answers whether premium models improve metrics.”

Eppo, which launched out of stealth in 2022, now has “several hundred” enterprise customers in its roster, including Twitch, SurveyMonkey, DraftKings, Coinbase, Descript and Perplexity, according to Sharma. Alexis Weill, Perplexity’s head of data, told TechCrunch that Eppo has allowed Perplexity to “significantly scale” the number of experiments it runs concurrently.

Investors seem pleased. This week, Eppo closed a $28 million Series B round led by Innovation Endeavors with participation from Icon Ventures, Amplify Partners and Menlo Ventures. Sharma says that the new cash, which values Eppo at $138 million post-money and brings its total raised to $47.5 million, will be put toward bolstering Eppo’s marketing and AI experimentation capabilities, enhancing its analytics offerings and scaling its go-to-market efforts.

San Francisco-based Eppo currently has 45 employees and expects to end the year with 65.

“The demands of efficient growth along with the rise of AI has combined to an adapt-or-die mentality that forces companies to become experimental,” Sharma said. “And due to the gaps of legacy vendors, most of the experimentation market had chosen to staff large in-house teams and build over buy. With so much employee movement and layoffs, these in-house teams are no longer sustainable, leading to companies seeking out Eppo to replace expensive or orphaned in-house builds.”

Read More

Defcon AI closes $44M seed round to solve a problem of ‘maximum complexity’: Military logistics

The U.S. Department of Defense is a mammoth organization. It not only employs millions of service members and hundreds of thousands of civilian employees, but also has the world’s largest military budget that’s used to buy and maintain more equipment than can likely fit into a single paragraph. 

It’s a lot to coordinate. Operators within the various agencies of the DOD must make decisions about how to plan their operations, coordinate resources and stay within budget for events that are likely contested — whether that’s from a hurricane or an adversary.

Two years after it was incubated, Virginia-based startup Defcon AI has raised a $44 million seed round led by Bessemer Venture Partners, with participation from Fifth Growth Fund and Red Cell Partners, among others, to solve this seemingly intractable problem.

Consider the Air Mobility Command, a command of the U.S. Air Force. When operators plan airlifts, they have to consider a whole slew of variables: available aircraft, the number of crews required, places for crews to rest, where to refuel, relevant airfields, cargo handling locations. Defcon AI says it has developed a set of software that allows the operator on the front end set these parameters “and then turn the software loose,” Defcon’s co-founder, chief strategy officer and retired U.S. Air Force General Paul Selva told TechCrunch. The software essentially operates against those parameters or inputs to generate the best plan — including the cost tables, resource requirements and schedule. 

This type of planning is difficult enough in the best of circumstances, but during a crisis, defense operators don’t even have the luxury of a day to allocate their resources. That’s where Defcon AI comes in.

“I’ve had all the jobs that we’re actually impacting,” Selva said. During his long military career, Selva held many titles, including the commander of the Air Mobility Command, which oversees nearly all of the Air Force’s fleet of air lift aircraft. He later became the commander of the U.S. Transportation Command, which coordinates transportation missions around the world, including those delivered by ships, trucks, trains and other forms of transportation. Before he retired in 2019, he was nominated by President Barack Obama to be the vice chairman of the Joint Chiefs of Staff. 

He co-founded Defcon in 2022 with Yisroel Brumer and Grant Verstandig, both founding partners of Red Cell Partners (Verstandig is also CEO). Red Cell has an interesting model: The firm makes internal investments but it also incubates companies (including Defcon), often identifying promising entrepreneurs that could lead them. Sometimes, entrepreneurs approach Red Cell before they found a company, and the firm handles things like board building, legal, HR and finance while the company grows. 

In the case of Defcon, Selva says that the company got started “because Air Mobility Command articulated a mission need that wasn’t being filled by industry.” The trio “had a conversation about whether or not we thought this was a tractable problem, and … our intuition was it is a mathematically and software tractable problem, but we have to do it a different way.” 

Brumer and Verstandig have their own impressive pedigrees. Before joining Red Cell, Brumer worked at the Pentagon as acting director of OSD/CAPE (Office of the Secretary of Defense, Cost Assessment and Program Evaluation), an enormous role that essentially functions as the “chief analytics officer” for the DOD, he said, and the overseer for the budget submission process. Verstandig is an entrepreneur who has incubated or grown businesses including Rally Health and defense startup Epirus. 

Defcon AI is targeting a problem of “maximal complexity,” Brumer said. The startup’s system combines different algorithms, including machine learning and mathematical optimization algorithms, to simulate a given scenario and generate the best logistical outcome to meet it. In the initial stages of product development, Defcon used reinforcement learning algorithms that don’t require data, but the company says it is now ingesting more and more data provided by the DOD to power the software. Operators can also choose whether to have the system simulate how an adversary might disrupt the operations, and can tell it to optimize for different variables, like speed versus cost effectiveness.

The company has earned around $15 million in government contracts and delivered a production version that was deployed for a real-world operation with Air Mobility Command less than two years after founding. The company is in the process right now of certifying the software to handle classified, secret information, both to expand its uses in the DOD and to enable it to ingest even more data. It’s also expanding to include trucks, trains and ships to its planning and simulation software.

Defcon is not planning on slowing down. The company sees even more applications across the DOD where its software can make an operational difference, and Brumer said they’re seeing “a very strong demand signal” from the private sector for the product too. Overall, the company says working closely with the end users will result in a better product and a genuine competitive edge in adversarial situations.

“Operational planners are actually trying to assess risk for their commanders,” Selva said. “They’re probably the most skeptical audience for decision support tools, so the extent to which you can partner with them you achieve a better outcome.” 

Read More

BeyondMath’s ‘digital wind tunnel’ puts a physics-based AI simulation to work on F1 cars

Simulating the real world is a tremendously complex problem if you want to do it at any useful level of fidelity. Traditional techniques are holding back design teams at vehicle and aerospace companies, but BeyondMath is putting AI on the task with a new way of simulating the world that could save them days or weeks of waiting.

“Unlike language, where we don’t have mathematical models to describe what the next word should be, when it comes to physics, we do have those models. And what we’re seeing is that machine learning is actually quite good at computation, not just pattern recognition,” said co-founder Darren Garvey.

The field in which BeyondMath is taking its first strides is called computational fluid dynamics (CFD), and it’s been around about as long as computing has. The equations that govern how an object moves through air or water, or air around an object, are fiendishly complex. So while we’ve continually improved our ability to predict, say, the way air flows over a wing, we’re still nowhere near perfect — and what we can do takes so much computational power that it’s limited to supercomputers and GPU clusters.

The result is that the design process in industries like cars, planes and boats involves a lot of wait time.

“For a designer, they put a lot of thought into what might work, then they run a simulation. Then they come in the next morning and they’ve got the results. Either it did what they wanted or not, and they have to go through this loop a few more times. Then you take it to the wind tunnel,” Garvey said — and the wind tunnel may well not agree with the simulation, so it’s back to the drawing board.

BeyondMath’s goal is to accelerate the digital design side, which means shortening the delay between having an idea and finding out whether it is likely to work.

“They’re saying, if I make this design change, will it make my car more fuel efficient? Imagine you’ve got six months to design a part for a plane. Given that a simulation takes so long, you might get 20 attempts to try things out. But if a designer thinks of an idea and gets results within seconds or a couple of minutes, in that same six months you might be able to run a million changes,” said Garvey.

Image Credits: BeyondMath

And it’s increasingly looking like machine learning, as opposed to just more GPUs running the same old equations, is the way to do that. Their first product is a “digital wind tunnel” that provides near-real-time simulation of airflow over a complex surface at a fidelity that would normally take hundreds of times as long.

We’ve seen something like this in scientific literature, where a model of a weather system can be effectively approximated in a fraction of the time, using a machine learning model trained on thousands of hours of simulations and observed patterns. But BeyondMath doesn’t have the luxury of a pre-existing training set.

“There’s just not a lot of simulation data out there — we don’t have the whole internet to train off of, like the LLMs. So how do you get something that’s equivalent to what designers are using, that works on these very complex geometries, as a startup?”

Surprisingly, the answer they’ve found is not to rely on simulations, but rather to have a model that understands the theory behind something like a wind tunnel, as well as the observed reality of that theory.

“We’re not trying to approximate the simulations, we’re trying to approximate the real world,” Garvey said. “And you have to bring in real-world data to do that.”

Once the model understands how a system behaves, it can also be an active participant in design, a possibility many engineers have already begun to explore in other domains. Garvey compared it to image understanding: There, too, machine learning models had to walk before they could run, but once they were adept at analyzing an image, it was an intuitive next step for them to generate one.

Among BeyondMath’s first markets is Formula 1 racing, where some unnamed teams are exploring using the software to speed up their aerodynamics and vehicle design processes.

“They’re one of the heaviest users of CFD, and they’re fast-moving, they’ll adopt new technologies. We’ve been working closely with a couple F1 teams, doing a lot of evaluation and understanding their core problems. We’re close to having a platform that will actually make their cars faster,” Garvey said.

The BeyondMath team (Garvey is second from the right).Image Credits: BeyondMath

In fact, he expressed hope (with the usual warning that there was no guarantee) that within six months “we’ll be able to show that customers are benefiting from these models, and they’ve gone out of research and proofs of concepts into things that have real impact.”

New funding should help make that happen: BeyondMath just raised an $8.5 million seed round led by UP.Partners, with Insight Partners and InMotion Ventures participating.

The startup expects to double its team size and scale up its compute; they’re buying Nvidia DGX 200s and working with the chip giant on this interesting new application of its ubiquitous compute hardware.

Though the highly competitive, deep-pocketed F1 racing community is certainly a good customer to have, BeyondMath is thinking about its next steps.

“We’re seeing a lot of success in our customers’ design space, but it’ll be a journey from that to something more generalizable. For example, if a model understands cars, or car-like objects, it’s not necessarily going to understand a plane, or a blood vessel,” Garvey said. “But that’s the classic startup dance — you have to find your path to traction before you have the runway to expand. As a business we’re focused on these top-tier customers so they can help bootstrap the company.”

Read More

Starpath accelerates moon water mining plans with $12M in funding

NASA and the space industry are in agreement: If we want to establish a permanent human presence on the moon, we’ll need to make use of every native resource we can — and none are as important as water ice. 

Starpath Robotics is one of a handful of startups planning for this future. The startup is betting that there will be a thriving market for liquid oxygen (LOX) refined from lunar water ice and that harvesting this resource will be key to humanity’s expansion throughout the solar system. LOX is a crucial component of propellant for some vehicles like rockets and spacecraft like lunar landers, and it’s used as the oxidizer alongside a combustible fuel such as hydrogen, kerosene, or methane.

The company came out of stealth last September with an ambitious water-harvesting architecture, involving mining rovers, refineries and LOX storage systems. Starpath wants to launch a demo mission “as quickly as we possibly can,” CEO Saurav Shroff said in a recent interview. To ensure the hardware will be ready whenever a launch vehicle is available, the company announced today that it has closed a $12 million seed round, co-led by 8VC and Fusion Fund, with participation from Day One Ventures, Balerion Space, and Indicator Ventures.

Much of the architecture has remained the same since Starpath originally unveiled its plans last year: Essentially, the company wants to use fleets of mining rovers that dig up hundreds of tons of lunar dirt and return it to autonomous lunar processing plans that extract the water, splits those molecules into their constituent atoms, and then liquifies the oxygen. The entire system would be powered by a massive solar array that’s being designed in collaboration with space solar startup Solestial.

There are some changes, however. Shroff said the 10-person team has made improvements on the mining rover hardware such that it will take far fewer rovers to produce a 1,000-ton annual harvesting scale (the company originally estimated it would take 50 rovers to hit that rate). Rover development also got a boost with $800,000 in NASA grants, as part of the space agency’s Break the Ice challenge to private industry. There is still much to develop, including the solar array and refineries, but the basic architecture is essentially planned out. With that in place, the company’s been turning its attention to the final interaction between its hardware and the customer’s vehicle. 

Figuring this out means solving a few problems, namely how to transport the LOX to the vehicle, which could be some distance away. The other problem, of course, is how to actually load the vehicle with the product. 

For the first issue, Shroff says the company is simply configuring its dirt-hauling rover to carry pressurized cryogenic propellant tanks instead. For the second, while he declined to go into specifics, he said the rover would be equipped with a specialized refueling instrument, that in all likelihood will be designed on a customer-by-customer basis. 

As of right now, there are few prospective customers for lunar LOX, but those that are planning moon missions could prove to be prolific buyers. Both SpaceX and Blue Origin have contracts with NASA to land on the moon before the end of the decade; Shroff estimates that Starship would consume around 100-300 tons of oxygen on the moon per flight, and something like Blue Origin’s Blue Moon would consume tens of tons of oxygen per flight. 

Given that Starpath is aiming to produce around 1,000 tons of LOX per year, regular Starship flights to and from the lunar surface alone could be enough demand to support this production capacity. At that rate, Shroff says, any operator would be able to fly a vehicle to the moon and trust they will be able to refuel it while they’re there. 

By the end of this year, the company is aiming to conduct an end-to-end, full-scale demonstration of its system in a simulated lunar environment, which they’ll build out at their 12,000-square-foot premises. After that, they’ll embark on a series of test campaigns before launching their first demonstration mission. The goal for that first mission is to harvest the equivalent of around 100 tons of liquid oxygen per year. It’s incredibly ambitious: Scientists have confirmed that water ice is on the moon, but no government or company has ever harvested it, let alone refined it in situ.

Starpath is currently at 10 full-time employees, and the new capital will primarily go toward doubling or even tripling that number at a fast pace. The payoff for getting this technology online could be enormous, the company is betting, with the moon just the first stepping stone to expansion through the solar system. Starpath already has its eyes on developing processing plans and rover harvesting fleets for Mars, which would be augmented from their lunar counterparts, to convert CO2 in the Martian atmosphere into methane.

“Life can be multiplanetary in a very short period of time,” Shroff said. “If you make 1,000 tons of liquid oxygen on the moon, your path to making a million-person city on Mars is hard, but it’s now possible.” 

Read More

Reliant’s paper-scouring AI takes on science’s data drudgery

AI models have proven capable of many things, but what tasks do we actually want them doing? Preferably drudgery — and there’s plenty of that in research and academia. Reliant hopes to specialize in the kind of time-consuming data extraction work that’s currently a specialty of tired grad students and interns.

“The best thing you can do with AI is improve the human experience: reduce menial labor and let people do the things that are important to them,” said CEO Karl Moritz Hermann. In the research world, where he and co-founders Marc Bellemare and Richard Schlegel have worked for years, literature review is one of the most common examples of this “menial labor.”

Every paper cites previous and related work, but finding these sources in the sea of science is not easy. And some, like systematic reviews, cite or use data from thousands.

For one study, Hermann recalled, “The authors had to look at 3,500 scientific publications, and a lot of them ended up not being relevant. It’s a ton of time spent extracting a tiny amount of useful information — this felt like something that really ought to be automated by AI.”

They knew that modern language models could do it: One experiment put ChatGPT on the task and found that it was able to extract data with an 11% error rate. Like many things LLMs can do, it’s impressive but nothing like what people actually need.

Image Credits: Reliant AI

“That’s just not good enough,” said Hermann. “For these knowledge tasks, menial as they may be, it’s very important that you don’t make mistakes.”

Reliant’s core product, Tabular, is based on an LLM in part (Llama 3.1), but augmented with other proprietary techniques, is considerably more effective. On the multi-thousand-study extraction above, they said it did the same task with zero errors.

What that means is you dump a thousand documents in, say you want this, that, and the other data out of them, and Reliant pores through them and finds that information — whether it’s perfectly labeled and structured or (far more likely) it isn’t. Then it pops all that data and any analyses you wanted done into a nice UI so you can dive down into individual cases.

“Our users need to be able to work with all the data all at once, and we’re building features to allow them to edit the data that’s there, or go from the data to the literature; we see our role as helping the users find where to spend their attention,” Hermann said.

Image Credits: Reliant

This tailored and effective application of AI — not as splashy as a digital friend but almost certainly much more viable — could accelerate science across a number of highly technical domains. Investors have taken note, funding an $11.3 million seed round; Tola Capital and Inovia Capital led the round, with angel Mike Volpi participating.

Like any application of AI, Reliant’s tech is very compute-intensive, which is why the company has bought its own hardware rather than renting it a la carte from one of the big providers. Going in-house with hardware offers both risk and reward: You have to make these expensive machines pay for themselves, but you get the chance to crack open the problem space with dedicated compute.

“One thing that we’ve found is it’s very challenging to give a good answer if you have limited time to give that answer,” Hermann explained — for instance, if a scientist asks the system to perform a novel extraction or analysis task on a hundred papers. It can be done quickly, or well, but not both — unless they predict what users might ask and figure out the answer, or something like it, ahead of time.

“The thing is, a lot of people have the same questions, so we can find the answers before they ask, as a starting point,” said Bellemare, the startup’s chief science officer. “We can distill 100 pages of text into something else, that may not be exactly what you want, but it’s easier for us to work with.”

Think about it this way: If you were going to extract the meaning from a thousand novels, would you wait until someone asked for the characters’ names to go through and grab them? Or would you just do that work ahead of time (along with things like locations, dates, relationships, etc.) knowing the data would likely be wanted? Certainly the latter — if you had the compute to spare.

This pre-extraction also gives the models time to resolve the inevitable ambiguities and assumptions found in different scientific domains. When one metric “indicates” another, it may not mean the same thing in pharmaceuticals as it does in pathology or clinical trials. Not only that, but language models tend to give different outputs depending on how they’re asked certain questions. So Reliant’s job has been to turn ambiguity into certainty — “and this is something you can only do if you’re willing to invest in a particular science or domain,” Hermann noted.

As a company, Reliant’s first focus is on establishing that the tech can pay for itself before attempting anything more ambitious. “In order to make interesting progress, you have to have a big vision but you also need to start with something concrete,” said Hermann. “From a startup survival point of view, we focus on for-profit companies, because they give us money to pay for our GPUs. We’re not selling this at a loss to customers.”

One might expect the firm to feel the heat from companies like OpenAI and Anthropic, which are pouring money into handling more structured tasks like database management and coding, or from implementation partners like Cohere and Scale. But Bellemare was optimistic: “We’re building this on a groundswell — any improvement in our tech stack is great for us. The LLM is one of maybe eight large machine learning models in there — the others are fully proprietary to us, made from scratch on data propriety to us.”

The transformation of the biotech and research industry into an AI-driven one is certainly only beginning and may be fairly patchwork for years to come. But Reliant seems to have found a strong footing to start from.

“If you want the 95% solution, and you just apologize profusely to one of your customers once in a while, great,” said Hermann. “We’re for where precision and recall really matter, and where mistakes really matter. And frankly, that’s enough; we’re happy to leave the rest to others.”

(This story originally had Hermann’s name incorrect — my own error, I have changed it throughout.)

Read More

Gemini Live could use some more rehearsals

What’s the point of chatting with a human-like bot if it’s an unreliable narrator — and has a colorless personality?

That’s the question I’ve been turning over in my head since I began testing Gemini Live, Google’s take on OpenAI’s Advanced Voice Mode, last week. Gemini Live is an attempt at a more engaging chatbot experience — one with realistic voices and the freedom to interrupt the bot at any point.

Gemini Live is “custom-tuned to be intuitive and have a back-and-forth, actual conversation,” Sissie Hsiao, GM for Gemini experiences at Google, told TechCrunch in May. “[It] can provide information more succinctly and answer more conversationally than, for example, if you’re interacting in just text. We think that an AI assistant should be able to solve complex problems … and also feel very natural and fluid when you engage with it.”

After spending a fair amount of time with Gemini Live, I can confirm that it is more free-flowing and natural-feeling than Google’s previous attempts at AI-powered voice interactions (see: Google Assistant). But it doesn’t address the problems of the underlying tech, like hallucinations and inconsistencies — and it introduces a few new ones.

The un-uncanny valley

Gemini Live is essentially a fancy text-to-speech engine bolted on top of Google’s latest generative AI models, Gemini 1.5 Pro and 1.5 Flash. The models generate text that the engine speaks aloud; a running transcript of conversations is a swipe away from the Gemini Live UI in the Gemini app on Android (and soon the Google app on iOS).

For the Gemini Live voice on my Pixel 8a, I chose Ursa, which Google describes as “mid-range” and “engaged.” (It sounded to me like a younger woman.) The company says it worked with professional actors to design Gemini Live’s 10 voices — and it shows. Ursa was indeed a step up in terms of its expressiveness from many of Google’s older synthetic voices, particularly the default Google Assistant voice.

But Ursa and the rest of the Gemini Live voices also maintain a dispassionate tone that steers far clear of uncanny valley territory. I’m not sure whether that’s intentional; users also can’t adjust the pitch, timbre or tenor of any of its voices, or even the pace at which the voice speaks, putting it at a distinct disadvantage to Advanced Voice Mode.

You won’t hear anything like Advanced Voice Mode’s laughing, breathing or shouting from Gemini Live either, or any hesitations or disfluencies (“ahs” and “uhms”). The chatbot keeps an even keel, coming across as a polite but apathetic assistant — as if Live has a multitude of conversations to handle and can’t invest particular attention to yours.

Chatting with Ursa

When Google unveiled Gemini Live at its I/O developer conference in May, it suggested that the feature could be useful for job interview prep. So I decided to give that a go first.

I told Gemini Live that I was applying for a tech journalism role, figuring I’d keep it simple and not step too far outside my area of expertise. The bot asked for details such as which specific job I might want within journalism (e.g. investigative versus breaking news reporting) and why, and then threw me a few generic practice questions (“Can you tell me a little about yourself?”) interspersed with more personalized ones (“What do you enjoy most about tech journalism?”).

I answered — a few sentences per question, nothing crazy — and asked Gemini Live for feedback. The chatbot was nothing if not complimentary. “Based on our practice, it sounds like you have a good handle on the challenges and rewards of tech journalism,” it said.

The glowing review seemed bit suspect to me, given my responses were completely off the cuff. So I tried throwing Gemini Live for a loop, disingenuously implying I’d only given one-word answers (I hadn’t) and asking the bot whether it was still sure about its critique.

It took the bait. “That’s true, you mostly gave one-word answers,” Gemini Live responded. “Can you tell me a little more about why you think you only gave one-word answers?”

This sort of gaslighting, if you will, was a common thread in my chats with Gemini Live. The model, like all generative AI models, confidently makes stuff up — a lot. And that makes it difficult to trust.

Strange behaviors

Gemini Live remembers a surprising number of things from earlier conversations in the same chat session; ask it to recount a story it told yesterday and there’s a reasonable chance it will. But when it comes to queries about people, places and things — queries you’d assume would be easier considering Gemini Live has Google Search at its disposal — the bot’s hallucinatory tendencies often rear their ugly heads.

Early on a Friday evening, I asked Gemini Live for suggestions of budget-friendly things to do in New York City. It recommended the infamous nightclub Le Bain (which it bizarrely pronounced “Le Boulebouleboulebou”), a rooftop bar called The Skylark and a second club, Output, in Williamsburg.

“Awesome!,” I thought… before I did some Googling.

Gemini Live had highlighted The Skylark’s supposedly “superb” happy hour, but I couldn’t find any mention of it on the bar’s website. Output permanently closed in 2019. Le Bain wasn’t a bad suggestion (it rarely charges cover) — but the line to get in is often an hour or more.

“You’re absolutely right! Output has indeed closed its doors,” Gemini said when I corrected it on the second point. “Thank you for catching that outdated information.”

I decided to switch gears, and had Gemini Live suggest a game to play.

“I’ll give you a word, and you have to create a new word using only the letters from the original word,” proposed the chatbot. That sounded appealing — until Gemini Live tried to convince me that the word “quiet” could be made from the letters in “cloud.”

Time for a new activity. Bored, I asked Gemini Live if it had any spicy takes.

“The increasing focus on mental health awareness is doing more harm than good,” Gemini responded. “By pathologizing normal human emotions and experiences, we’re creating a generation overly reliant on medication and therapy, rather than fostering resilience and coping mechanisms. Boom!“

Note: This is a screenshot of the chat log for Gemini Live, not the Gemini Live UI.Image Credits: Google

Spicy to be sure — and controversial!

“That was a provocative statement designed to spark discussion rather than a well-rounded analysis of a complex issue,” Gemini Live told me when I asked whether it was sure about what it said. “The increasing focus on mental health awareness is undoubtedly a positive development.”

Wishy-washy

Gemini Live’s dueling takes on mental health illustrate how exasperatingly nonspecific the bot can be. Even where its responses appear to be grounded in fact, they’re generic to the point that they’re not incredibly useful.

Take, for example, my job interview feedback. Gemini Live recommended that I “focus my interview prep” and “practice talking about my passion for the industry.” But even after I asked for more detailed notes with specific references to my answers, Gemini stuck to the sort of broad advice you might hear at a college career fair — e.g. “elaborate on your thoughts” and “spin challenges into positives.”

Where the questions concerned current events, like the ongoing war in Gaza and the recent Google Search antitrust decision, I found Gemini Live to be mostly correct — albeit long-winded and overly wordy. Answers that could’ve been a paragraph were lecture-length, and I found myself having to interrupt the bot to stop it from droning on. And on. And on.

Image Credits: Google

Some content Gemini Live refused to respond to altogether, however. I read it Congresswoman Nancy Pelosi’s criticism of California’s proposed AI bill SB 1047, and, about midway through, the bot interrupted me and said that it “couldn’t comment on elections and political figures.” (Gemini Live isn’t coming for political speechwriters’ jobs just yet, it seems.)

Image Credits: Google

I had no qualms interrupting Gemini back. But on the subject, I do think that there’s work to be done to make interjecting in conversations with it feel less awkward. The way it happens now is, Gemini Live quiets its voice but continues talking when it detects someone might be speaking. This is discombobulating — it’s tough to keep your thoughts straight with Gemini chattering away — and especially irritating when there’s a misfire, like when Gemini picks up noise in the background.

In search of purpose

I’d be remiss if I didn’t mention Gemini Live’s many technical issues.

Getting it to work in the first place was a chore. Gemini Live only activated for me after I followed the steps in this Reddit thread — steps that aren’t particularly intuitive and really shouldn’t be necessary in the first place.

During our chats, Gemini Live’s voice would inexplicably cut out a few words into a response. Asking it to repeat itself helped, but it could take several tries before the chatbot would spit out the answer in its entirety. Other times, Gemini Live wouldn’t “hear” my response the first go-around. I’d have to tap the “Pause” button in the Gemini Live UI repeatedly to get the bot to recognize that I’d said something.

This isn’t so much a bug as an oversight, but I’ll note here that Gemini Live doesn’t support many of the integrations that Google’s text-based Gemini chatbot does (at least not yet). That means you can’t, for example, ask it to summarize emails in your Gmail inbox or queue up a playlist on YouTube Music.

So we’re left with a bare-bones bot that can’t be trusted to get things right and, frankly, is a humdrum conversation partner.

After spending several days using it, I’m not sure what exactly Gemini Live’s good for — especially considering it’s exclusive to Google’s $20-per-month Google One AI Premium Plan. Perhaps the real utility will come once Live can interpret images and real-time video, which Google says will arrive in an update later this year.

But this version feels like a prototype. Lacking the expressiveness of Advanced Voice Mode (to be fair, there’s debate as to whether that expressiveness is a positive thing), there’s not much reason to use Gemini Live over the text-based Gemini experience. In fact, I’d argue that the text-based Gemini is more useful at the moment. And that doesn’t reflect well on Live at all.

Gemini Live wasn’t a fan of mine either.

“You directly challenged my statements or questions without providing further context or explanation,” the bot said when I asked it to scrutinize my interactions with it. “Your responses were often brief and lacked elaboration [and] you frequently shifted the conversation abruptly, making it difficult to maintain a coherent dialogue.”

Image Credits: Google

Fair enough, Gemini Live. Fair enough.

Read More

Now a million people can watch you fumble Zoom’s screen-share settings at once

Zoom on Monday announced a new single-user webinar feature that caps out at 1 million attendees. The addition comes less than a month after the #WinWithBlackWomen fundraiser for Vice President Kamala Harris peaked at 44,000 users before crashing. More recently, an X interview between owner Elon Musk and Harris’ challenger, Donald Trump, crashed prior to launch.  

Given the audience capabilities of Zoom’s new feature — which also comes in 10,000, 50,000, 100,000, 250,000 and 500,000 flavors — the company is offering up live support from its Event Services. The Zoom reps will “ensure hosts deliver a professional, engaging experience” and, presumably, make sure it doesn’t crash under its own weight.

Read More

Could Trump’s AI-generated Taylor Swift endorsement be illegal?

On Sunday, former President Donald Trump posted a collection of memes on Truth Social — the platform owned by his media company — that make it seem like Taylor Swift and her fans are coming out in support of his candidacy. But as new legislation takes effect, these images could have deeper implications about the use of AI-generated images in political campaigns, especially when those images misrepresent a celebrity’s likeness.

“One of the things I’m seeing a lot of in my practice right now is the rise of AI impersonators across the board for endorsements,” Noah Downs, an IP and entertainment lawyer, told TechCrunch, with the condition that his comments are not legal advice. These fake AI endorsements have become so widespread that even “Shark Tank” had to publish a PSA to warn fans about the prevalence of scams that impersonate the show’s investors.

In one of the images Trump posted, hordes of young women wear matching “Swifties for Trump” t-shirts. While there is indeed political diversity among the large population of Swift fans, these images appear to be AI-generated — in fact, these particular images come from a satirical post on X.

Another meme that Trump posted is a screenshot from X that depicts Taylor Swift as Uncle Sam, declaring, “Taylor wants you to vote for Donald Trump.”

Image Credits: Screenshot posted by Donald Trump on Truth Social

Though the pop icon has not yet commented on the 2024 U.S. presidential election, she came out in support of the Biden-Harris campaign in 2020 and publicly disparaged Trump at the time. Some fans speculated that Swift had subtly endorsed Harris in an Instagram post this month, though this was not the case.

As one of the most dominant figures in pop culture, Swift has been subject to her fair share of deepfakes. When non-consensual, explicit AI images depicting Swift went viral on X this year, some lawmakers responded by introducing new bills aiming to protect against deepfakes. Even White House Press Secretary Karine Jean-Pierre called on Congress to do something.

Eight months later, the landscape of legal protections against misleading synthetic media already looks different. In Tennessee, where Swift’s corporate representation is based, Governor Bill Lee in March signed the trailblazing ELVIS Act into law, which carves out explicit protections for artists against unauthorized AI imitations of their work.

“This legislation was passed with bipartisan support, because everyone appears to recognize the problems that AI and misuse of AI tools can present to the public,” Downs said.

But since the ELVIS Act is so new, there isn’t precedent for how it could be used to protect artists. Much of the legislation’s language focuses specifically on AI-generated audio that can mimic an artist’s voice, like the viral Drake song that turned out to be fake.

“I do think that this is going to be a long-term issue that the ELVIS Act is very prescient in taking care of, but we need to have more robust national legislation about it,” Downs said. The only reason why the ELVIS Act may even potentially be at play is because of Swift’s connections to the state, where she has business and real estate holdings.

Avi D. Kelin, a partner at PEM Law focusing on political law, is not optimistic that the ELVIS Act could apply, since the law appears more concerned with audio-based impersonation than imagery. Instead, he wonders if this could become a federal election integrity concern in the future.

“The larger question is whether the Federal Election Commission, which has jurisdiction over political communications, will get involved,” Kelin told TechCrunch. But he said that the FEC doesn’t seem likely to roll out new guidelines on AI-generated political communications this election cycle.

The Federal Communications Commission (FCC), however, announced it is moving forward with plans to enact new AI transparency requirements on TV and radio advertisements. But that doesn’t apply to social media posts by politicians running for government office, and social media remains a key component of campaign communications. Meanwhile, research from the Center for Countering Digital Hate (CCDH), a British nonprofit focused on online extremism, showed that the volume of AI-generated disinformation increased an average of 130% per month on X over the last year.

These disingenuous endorsements matter so much because Swift’s support is perhaps the most coveted celebrity backing a politician can get. Her cultural influence is so vast that her support of a candidate could tip the scales in a tight race — according to Morning Consult, more than half of adults in the U.S. consider themselves fans of Taylor Swift, while 16% identify as avid fans. Those numbers are staggering given the context that only about two-thirds of eligible Americans cast their vote in the 2020 election.

“The [ELVIS Act] is brand new, and the exact parameters will need to be developed by the courts,” said Kelin. “This would certainly be an interesting test case!”

Read More

From a $2.5 million hyper car to a Spanish track-ready EV, here were the most interesting EVs at Monterey Car Week

Historic vehicles, flowing champagne and fashion have dominated the events at Monterey Car Week for decades now. But a change is afoot: EVs, tech-centric vehicles, startups and a heavy dose of Silicon Valley’s software developers and founders added to the scenery this year. 

Even the Pebble Beach Concours d’Elegance seemed more youthful and modern — from the crowd to the cars — a departure for the Sunday car show that closes out the week of automotive festivities. The Concours d’ Elegance, the spot to revel in pre-war and interwar vehicles, took the unusual step of highlighting Wedge-style cars from the 1950s through 2023. One of them — a 1970 Lancia Stratos HF Zero Bertone Coupe owned by Phillip Sarofim, the founder and CEO of Trousdale Ventures and chairman of Meyers Manx — even made it to the top three finishers for best in show. The top honors ended up going to a 1934 Bugatti Type 59 Sports.

A 1970 Lancia Stratos HF Zero Bertone Coupe at Pebble Beach Concours d’Elegance on August 18, 2024 in Monterey, California.Image Credits: Matt Jelonek / Getty Images

Amid the crowds of collectors, wannabe social media influencers, young car fans and a small handful of celebrities, a combination of heritage brands and tiny startups made waves with the wealthy who descended to see some of the rarest and most expensive vehicles in the world. 

Most notably, the crowd at this year’s events was both larger and younger. At events like Hagerty’s Motorlux on Wednesday night, and Friday’s The Quail, A Motorsports Gathering, tech founders made their rounds, drinking champagne, enjoying walls of bacon and limitless amounts of caviar on crackers, and even dancing the night away — a rarity at the events that are more likely to be punctuated by classical music than dance music. 

Rob Howard, founder of Bay Area-based Kindred Motorworks, said the younger crowds were apparent, lively and clearly wanted to experience — not just look at — the cars. Another founder in his mid 40s, who asked not to be named, complained he was the oldest person at the McLaren party.

And parties — there were plenty. 

More than 4,100 people attended the Motorlux event at the Monterey Jet Center, according to Hagerty public relations, where attendees spent an hour or more in traffic just to find parking. The Quail boasted more than 5,000 attendees on Friday. There were fewer unauthorized road rallies, thanks to a very robust, and very visible, police presence in and around Monterey, Carmel and Pebble Beach, and far more excitement around electrification than there has been in the past. 

Howard also noted that while EVs have been a part of Monterey Car Week before, this year they seemed to seep in a bit deeper. His startup, which unveiled a vintage (and now electric) Bronco at Hagerty’s Motorlux event, was part of that trend. 

Here’s a quick roundup of the best things we saw at the annual auto-enthusiasts’ bacchanal. 

Acura Performance EV Concept

Acura Performance EV ConceptImage Credits: Abigail Bassett

Acura, the luxury arm of the Japanese brand Honda, showed off its Performance EV Concept at The Quail on Friday. While most concepts never see the light of day, the Performance EV Concept represents an upcoming model that will go into production next year and be built at Honda’s new EV hub in Ohio. The fully electric underpinnings of this coupe-uv (as in utility vehicle) will be built by Honda, unlike the current Acura ZDX (and the Honda Prologue), which sits on GM’s Ultium platform.

BMW M5 Touring

BMW M5 wagonImage Credits: Abigail Bassett

BMW debuted an M5 wagon at Pebble Beach this year and made automotive enthusiasts very happy, since it marks the first time that BMW has decided to sell an M longroof stateside.

The most interesting part is the powertrain that underpins the 717-horsepower sleeper: Both the new M5 sedan and the wagon get a plug-in hybrid powertrain loosely based on the one in the M8 hybrid race car. Both the M5 sedan and M5 Touring (the wagon) get a 14.8 kW lithium-ion battery which powers a 194-hp motor integrated into the eight-speed transmission, and all married to a 4.4-liter twin-turbo V8 engine. It remains to be seen how frequently an M5 buyer will actually plug in to recharge. 

Everatti 

Everatti Land Rover IIAImage Credits: Abigail Bassett

Everatti takes shells of old cars, like a 1960s-era Mercedes-Benz 280 SL (known as the Pagoda), scans them, restores them and drops an EV-powertrain into them, while keeping the same ride quality, power and more. 

The U.K.-based startup showed off (and let us drive) an electric right-hand-drive Land Rover IIA on the roads around the Lodge at Pebble Beach, and it was a joy. The company managed to capture the character of the old Land Rover while making it completely electric. Everatti has created a Porsche 911 (969) RSR as well as the Pagoda, working with individual buyers to turn classics into EV-powered masterpieces. 

These don’t come cheap though. Prices start around $225,000, and you’ll either have to find a donor car, or the company can help you do so. And if you’re a combustion engine purist, worried about matching serial numbers, never fear, Everatti hangs onto all the components should you ever want to convert the vehicle back to its original fuel-burning version. 

Hispano Suiza Carmen Sagrera

Hispano Suiza Carmen SagreraImage Credits: Abigail Bassett

Hispano Suiza isn’t necessarily a household name, but the 120-year-old Spanish coachworks builder showed off its track-focused all-electric Carmen Sagrera, which makes a whopping 1,114 horsepower and 848 lb-ft of torque, powered by a 103 kW battery pack at The Quail. The company has been owned by the same Suqué Mateu family since 1904. Only 24 Sagreras will be made. 

Karma Kaveya

Interior of the Karma KaveyaImage Credits: Abigail Bassett

Karma Automotive — not to be confused with the vehicle from the defunct automaker, Fisker — showed off its Kaveya, an EV with 1,000 horsepower, at The Quail on Friday. 

The Karma Kaveya super coupe is the first vehicle developed by the California automaker following a deal the company has with Intel Automotive. While the exterior of the “software-defined vehicle architecture’’ (SDVA) is still in flux, the interior is “production ready,” with both companies stating that they plan to “benchmark, demonstrate and validate critical concepts for the advancement of open standards for SDVA that can be shared openly and commercially to support the transformation of the broader automotive industry.” 

There’s a whiff of vaporware around the company, but we’ll have to wait and see how things shake out. 

Kindred Motorworks EV Bronco

Kindred Motorworks EV BroncoImage Credits: Abigail Bassett

Kindred Motorworks brought its first all-electric Bronco to the events in and around Pebble Beach, and according to Howard, it brought considerable attention to the four-year-old restomod company. Kindred showed off an EV restomod, a restored 1960s-era Bronco, powered by two motors and an 80kWh battery pack at Motorlux, and shuttled attendees to the events at The Quail Lodge on Friday. The Bronco EV is the first fully electric Kindred vehicle that will be delivered to customers.

Lamborghini Temerario

Lamborghini TemerarioImage Credits: Abigail Bassett

No Pebble event is complete until you’ve had more than one conversation (rudely) interrupted by a roaring Lamborghini. The company screamed onto the scene at The Quail, A Motorsports Gathering with their successor to the Huracán; a plug-in hybrid Lambo sports car called the Temerario. 

Rather than that raucous V10 that we’ve all come to, uh, recognize, the Temerario gets a V8 married to a very tiny 3.8 kW battery mounted down the center of the car to keep weight balance more even. That battery powers three motors mounted front and rear. Two mounted on the front axle give the Temerario the equivalent of an electric all-wheel drive system. The third is mounted between the engine and the eight-speed, dual-clutch transmission and helps manage torque between shifts.

The Temerario makes a peak combined horsepower of 907 hp, and Lamborghini has yet to certify how much electric range the new vehicle will get, but it should be on par with the Revuelto. Like most high-end automakers, Lamborghini continues to cling to its internal combustion engine history despite industry pressures to electrify.

RBW Roadster and GT

RBW Roadster Image Credits: Abigail Bassett

While these two vehicles look like old MGs, underneath they’re all EV. They get an estimated 150 miles of range from a small 22kw battery nestled into the frame and come complete with all the modern amenities you could wish for, including Bluetooth, Apple CarPlay, crash sensors and air conditioning. 

RBW is a U.K.-based startup and they’ve been doing a lot of work behind the scenes to get these adorable, tiny, British-looking cars to be road legal both stateside and overseas. The company only plans to make just shy of 100 of these a year, but they have plans for a U.S.-based production location coming soon, which will help them scale. Colors and interiors are fully customizable and the Roadster is $139,0000 while the GT is $151,000.

Rimac Nevera R

Rimac Nevera R Image Credits: Abigail Bassett

The Croatian startup that merged with Bugatti wowed the crowd at The Quail this year with their juiced-up Nevera R — a hyper car with even more power than the Nevera. The R version makes 2,107 horsepower and does zero to 60 in less than 2 seconds.

Rimac promised more to come at future Quail events, teasing a “track-focused car that’s faster than any other track car.” The CEO told the gathered crowd that the company has “some work to do on the racetracks around the world before we show it here at Quail.” The Nevera R will be priced at around $2.5 million, and only 40 will be made. 

Porsche 993 Speedster

Porsche 993 Speedster Image Credits: Abigail Bassett

For a particular price, Porsche will work with you to create the car of your dreams, thanks to their Sonderwunsch (Special Request) division — and build vehicles like this “Otto” yellow speedster (a color named after the owner, Luca Trazzi’s beloved dog) with the underpinnings of a 993 Carrera RS, enmeshed with a 1994 Porsche 911 Cabriolet, and a 993 Turbo.

Trazzi is a designer with a penchant for Speedsters and a robust collection of Porsches. According to reporting at Road & Track, he worked so closely with the Sonderwunsch group on the one-off project that he even got his own Porsche ID badge.

Read More